Instruction-tuned Language Models are Better Knowledge Learners